Castle Labs Review

How Castle Labs scores in incentive program operators

Castle Labs is rated in our incentive and grant program operators ranking in the program evaluation role. Its scored work is the assessment of incentive programs it wrote with DefiLlama Research as Arbitrum DAO's research member, posting on the Arbitrum forum as CastleCapital, plus a retention study published on its own research site. The review is desk research of those pages, read on November 6, 2026.

The six criteria use fixed scales set out in our ratings methodology. For an evaluator, the programs it assessed count as its named programs.

The overall score is the plain average of the six marks: (10 + 5 + 10 + 5 + 5 + 5) / 6 = 40 / 6 = 6.7.

One mark changed at this check, and the overall fell from 7.5 to 6.7. Disclosed volume fell from 10 to 5. Under the category's rule, budgets do not count as distributed volume, and the assessment labels most of its per-program figures as budgets. The earlier review had counted them as amounts distributed. Where the report does give paid-out amounts, as for STIP and zkSync Ignite, they now earn the partial mark.

Strongest criteria: named programs and impact report

Named programs (10). The rubric's top mark needs five or more programs, each with its network and period. The assessment covers 11 programs on four networks, four of them on Arbitrum with start and end months. Two Avalanche programs are listed with "no specific timeline", but the remaining programs are well above the five needed.

Impact report (10). This is the criterion where an evaluator should be strongest, and Castle publishes both parts the scale asks for: metrics and how they were measured. The retention study fixes its definitions before giving results. A user is a supplier of interest-bearing capital, success is deposit volume, and monthly cohorts of new depositors are tracked. The results are not flattering to the program: deposits relative to each cohort's first month were lower for incentive-period cohorts than for earlier ones. Our guide to yield farming explains why reward-driven deposits tend to leave.

Weakest criteria: four marks of 5

All four 5s are partial disclosure. Castle publishes something on each, but not enough for the top mark.

• Disclosed volume. Budgets are stated for every program, but a budget is a plan, not a payout. The zkSync Ignite case shows the gap: the planned 325 million ZK became 45 million distributed.
• Abuse protection. Sybil resistance appears as an evaluation question and a recommendation. Castle describes no detection method of its own.
• Open data. One public Dune dashboard backs the retention study. The main assessment's underlying data is not published as a set.
• Independence. The disclosure names two ties but not the protocols Castle works for. That matters here: its homepage, which sells research, distribution and advisory services, lists Arbitrum under "Trusted by", and Castle also evaluates Arbitrum's programs. No relationship list is set against that work.

What the score does not cover

The rubric scores what Castle publishes about the programs it evaluated. It does not judge whether its recommendations were right, and it does not audit the figures it reports. It also does not cover Castle's paid research and advisory work, its grants-allocator decisions, or its delegate votes, except where they bear on the independence criterion.

Castle Labs vs the closest rated peers

At 6.7, Castle Labs ties Gitcoin and Metrom at the time of this check. Ties are listed alphabetically, so Castle Labs comes first. OpenBlock Labs, which the assessment credits for fraud detection on two programs, scores 6.3, as do Questbook and Thrive Protocol. Merkl scores 5.8. Karma, another operator working on grant accountability, scores 5.0.

Who it suits

The score suits DAO delegates and program designers who want an outside reading of past incentive programs with a documented method, especially on whether users stayed after the rewards ended. Readers should keep the open independence question in mind when an assessment covers a protocol Castle may also work for.